DOI: https://doi.org/10.36347/sjpms.2026.v13i04.001

Pages: 150-165

We present a sensorless thermometry method that reconstructs perovskite module temperature directly from routine MPPT logs, requiring no paid instrumentation. The approach separates a fast irradiance-driven term from a slow thermal response modeled as a single-RC state. A two-regressor, sliding-window identification is solved with orthogonalization and mild ridge regularization; temperature is then recovered by physics-informed inversion, with residual bootstrap providing confidence intervals and automated gates preventing ill-posed Kelvin reporting. Using perovskite-like priors (temperature-coefficient magnitude 1.3" "×10^(-3)–2.0×10^(-3) " " K^(-1) and thermal time constant 50–400 s), 45–60 min logs at 1 Hz consistently recover the trajectory shape of T(t). Absolute-Kelvin accuracy without any anchor depends on excitation and τ: median MAE spans ∼5–7 K at τ=50s and rises to ∼25–27 K at τ=240s and ∼43K at τ=400s in high-variability records, while the relative error against the temperature swing remains comparatively stable. Uncertainty bands contract during strongly forced segments and widen during quiescent intervals, matching identifiability diagnostics. In practice, one brief temperature anchor (ambient or back-of-module) fixes the absolute scale and tightens Kelvin-level errors; including V_mpp (t) and I_mpp (t) further improves regressor separability. The fully reproducible Python pipeline (NumPy/pandas/pvlib-python) aligns with low-cost, ambient workflows emphasized for perovskite research.